Fast Shadow Detection from a Single Image Using a Patched CNN

This code is for the paper: S Hosseinzadeh, etc. “Fast Shadow Detection from a Single Image Using a Patched Convolutional Neural Network”, Proceedings of the IEEE/IROS 2018, https://arxiv.org/abs/1709.09283.

Generating The Shadow Prior Map Images

• These images are used as image-level prior that are defined in line 44 of main_fast_shadow_detection.py.
• Install Paired_Region_Prob_Map using README in the folder. Run Paired_Region_Prob_Map/deshadow_driver.m by MATLAB
• Reference paper http://dhoiem.cs.illinois.edu/publications/pami12_shadow.pdf

Dependencies:

1. nolearn
2. lasagne
3. theano
4. scipy
5. sklearn
6. matplotlib
7. skimage
8. Python’s basic libraries (pickle, sys, os, urllib, gzip, cPickle, h5py, math, time, pdb)

How To Run The Code:

• python2.7: run main_fast_shadow_detection.py
• python3: run main_fast_shadow_detection_p3.py

Notes:

• Build folders “data_cache” and “prediction_output_v1” for data training/testing output files, and output prediction result files.
• TrainImgeFolder: Training Images
• TrainMaskFolder: Training Masks (Ground Truth)
• TrainFCNFolder: Shadow Prior Map Images

• Likewise for testing images…
• The Mask and Shadow Prior files should have 1 dimension, and Mask files also should be binary.

Using GPU:

Build a file in your home ~/.theanorc with a content of:

[global]
floatX = float32
[nvcc]
fastmath = True